Skip to content
BSWEN
Blog

Claude skills vs mcp

Cowrie
Cowrie
Dev @ Bswen

Introduction

In this post, I will compare the brand new claude skills and mcp(model context protocol) ,both are created by anthropic.

what is mcp

The Model Context Protocol (MCP) is an open standard that allows AI agent to connect to external data, applications, and services.

Think of it as a “USB hub” for AI agent, providing a standardized way for AI agents to perform actions, It replaces the need for custom integrations for every new data source.

Just as the following picture shows, the AI agent utilize the mcp protocol to connect to datasources, execute commands … , it make AI agent more useful for human. ==It can impacts something in the world, not just talking.

MCP make the AI agent more actionable, because it can connect to different types of mcp servers using the same protocol.

what is claude skills

Skill is a persistent knowledge to complete a task for AI models. The complete definition of skills is :

It’s a "folder/module" that contains: SKILL.md (a Markdown file with explanations, including metadata and instructions), optional scripts/resources, and potential templates or tools.

A skill is just a folder or zip file, it contains two main parts, one is SKILL.md, a markdown file describe what this skill is and how to use it to complete a task, this file also contains content to help AI agent to understand the background and scenerio , make AI more intelligent to use this skill.

Why Launch Skills Despite Existing MCP?

Because there are some problems when using MCP in an AI agent:

  • Low accuracy and high token consumption., to complete a simple task, AI agent need to consume a lot of tokens to interact with mcp servers.

  • Complex and high learning curve: Using MCP requires specialized knowledge, setting up MCP servers, and understanding its communication protocols (like stdio, SSE, etc.). And sometimes, you need to know what is uvx and pip, and need to know json.

  • Only provide a tool: It doesn’t tell AI how to combine tools to solve a real problem in real world — relies on prompt engineering, which is hard to reproduce.

  • Protocol-oriented, not task-oriented: MCP knows how to make one dish, but not how to arrange a full banquet.

The following picture shows the main problem of MCP, it only tell AI there is tool named xxxx, and it can do something, but in real world, we need to solve problems, for example, provide a banquet, then it comes skills, it tell AI how to make use of different tools to complete a real job.

The relationship between skill and mcp

Skill is more task-oriented and mcp is tool-oriented, skill can use mcp to do a task. Skill tell AI agent to complete a task by workflow, which is constructed by scripts/resources or mcp servers.

Here is a diagram shows their relationship:

  • AI agent read skill
  • The skill contains guide to use mcp tools to complete a task
  • AI agent follows the guide and call the mcp tool to complete a task

You can see that the skill is optional to AI Agent, it can just call mcp tools directly without skill, but it will need more prompts to tell it how to use these tools to do a job.

But AI agent will consume more tokens to do the job, and the knowledge prompts are not portable , others can not reuse the knowledge conveniently to solve same problems.

What makes skill different from mcp?

First , skill is composable, one skill can depends on another skill: But there is no depends for mcp servers for now, if you want to achieve the same effects, you must use mcp proxy server or do some tricky on mcp client side.

Second, skill is portable, because essentially a skill is just a zip file, you can copy and paste everywhere , although the mcp server is also portable, but it does not contain knowledge to do a real job, it’s just a tool.

Why skill consumes less tokens

First, skill supports progressive loading, e.g. AI agent reads the entry SKILL.md, and if it loads more resource files by user task need.

But for mcp, AI agent needs to load all the mcp servers ,and then make a decision to choose some mcp server to do the job.

Summary

I think skill is a more advanced prompt engineering technique than mcp servers, it contains real knowledge to solve real problems in real world. Though it is now a dedicated in anthropic claude tools, I think it will be accepted by AI industry.

Final Words + More Resources

My intention with this article was to help others share my knowledge and experience. If you want to contact me, you can contact by email: Email me

Here are also the most important links from this article along with some further resources that will help you in this scope:

Oh, and if you found these resources useful, don’t forget to support me by starring the repo on GitHub!

Tags: